Joint Vehicle Selection and Resource Allocation for Hierarchical Federated Learning in Internet of Vehicles

Abstract

Hierarchical federated learning (HFL) in the Internet of Vehicles (IoV) faces challenges in participation reliability and aggregation efficiency due to vehicle mobility and resource heterogeneity. This paper investigates joint vehicle selection and resource allocation in a three-tier HFL architecture. In this framework, roadside units (RSUs) perform synchronous edge aggregation, while a macro base station (MBS) conducts asynchronous global aggregation. We derive a participation probability based on mobility, local computing capability, and statistical channel state information to construct the eligible vehicle set before final vehicle selection and channel allocation. Furthermore, we design an age-of-model (AoM)-aware policy to mitigate the impact of stale edge models. To maximize HFL training efficiency, we model the system as a multi-agent Markov decision process and propose a multi-agent discrete soft actor–critic (MA-DSAC) algorithm. Simulations on CIFAR-10 and Fashion-MNIST verify that MA-DSAC improves delay-constrained convergence, reducing the delay required to reach the target accuracy by up to 70% and maintaining a lower average AoM than the considered baselines.